Speaker
Description
In the study of the Universe, particularly galactic dynamics,
hydrodynamical simulations have played a critical role. Finely-tuned on a wide range of scales, recent large-scale simulation suites have produced mock observations that are nearly indistinguishable from actual observations.
In comparison to traditional machine learning techniques, the
complexity of these models poses a substantial challenge. However, artificial neural networks have the ability to learn by transforming data iteratively into a more compressed and abstract representation. Once optimized, these networks can generalize to unseen data, which makes them more applicable in a "real-world" setting.
In this talk, we would like to provide a status update on the current effort at ZHAW to explore the mapping from cosmological simulations to mock images from simulated SKA visibilities (and vice versa), using deep learning techniques. We will present a comparison of the performance of two image-to-image translation conditional adversarial network models, pix2pix and CycleGAN, trained on a dataset of galaxies from the IllustrisTNG simulation suite.
